Deep-learning based estimation of loco-regional control for patients with locally advanced HNSCC

Purpose/Objective: In order to improve radiotherapy outcomes, further treatment personalisation is considered beneficial. Radiomics analyses aim to predict treatment outcomes based on medical imaging data. Commonly, hand-crafted imaging features are used that require domain knowledge and further feature selection steps. This may cause relevant information to be lost. Deep convolutional neural networks (CNNs) on the other hand can act as automatic feature detectors and are able to learn highly nonlinear relationships directly from imaging data, thus addressing the drawbacks of conventional radiomics approaches and enabling end-to-end learning. We investigated whether CNNs are capable of quantifying loco-regional tumour control (LRC) based on CT imaging of patients with locally advanced head and neck squamous cell carcinoma (HNSCC).
Material/Methods: A multicentre cohort consisting of 302 patients with locally advanced HNSCC was collected and divided into an exploratory and a validation cohort (207 and 95 patients, respectively). All patients received a non-contrast-enhanced CT scan for treatment-planning and were treated by primary radio(chemo)therapy. 9725 transverse CT slices from the exploratory cohort were used to train a CNN with eight convolutional layers. For every patient (with one exception) we used 23 CT slices cranial and caudal of the slice with the largest tumour area, resulting in 47 slices per patient. Discriminative performance was evaluated using 4465 slices of the validation data set. The hazard of loco-regional recurrence was estimated by the CNN maximising the likelihood of the Cox proportional hazards model, which allows for incorporation of nonlinear relationships between the imaging features and the hazard prediction. The final hazard for every patient was obtained by averaging the results of the individual slices. The prognostic value of the model was evaluated by the concordance index (C-Index). Patients were stratified into groups of low and high risk of recurrence using the median hazard in the exploratory cohort.
Results: The validation of our CNN model revealed a C-Index of 0.68 (95% confidence interval: 0.57-0.79) for the prognosis of LRC. The estimated hazards were used to stratify patients into two risk groups. LRC significantly differed between these groups, both in the exploratory and the validation cohort (log-rank p<0.0001 and p=0.0005, respectively). Compared to previously published results with an average validation C-Index of 0.62 based on conventional radiomics [1], prognostic performance was slightly improved.
Conclusions: We showed that CNNs are capable of automatically stratifying patients with locally advanced HNSCC into high and low-risk groups for loco-regional tumour recurrence. The obtained results suggest that deep-learning based approaches can become useful for non-invasively evaluating individual recurrence risks encouraging future research in this area.
[1] Leger et al. Sci Rep 7: 13206 (2017).